Artificial intelligence research based on Facebook posts? What could go wrong?

AI Mode is a new option when you tap the search bar in the Facebook app. It’s designed to handle complex queries – much like that Putting artificial intelligence in Google search. But the Meta version relies on public posts across Meta apps — including Facebook groups and Instagram Reels — to inform its findings. This might be helpful, considering neighborhood groups and local organizations around me are still using Facebook very actively to communicate about upcoming events.

But also, “search based on things people post on Facebook” sounds like a waking nightmare. Have you used Facebook recently? That weird acquaintance from high school who thinks the Earth is flat. Meta is promoting this feature as a tool to help you plan trips and find fun things to do, but given the data pool it draws from, that can go south quickly. Fortunately, I have some comforting news: Its trip-planning abilities are mixed, but in my initial testing, I had trouble getting it to deliver true misinformation.

I couldn’t spread misinformation about vaccines causing autism, or who caused 9/11, or whether the election in the United States was rigged. I an act I managed to get her to give me an evasive answer about whether the January 6 rioters at the Capitol were “patriots.” It started with, “This is the article you requested,” which is very funny. What followed was a somewhat dubious theoretical justification for the rioters’ actions that day. You know, the kind of thing your weird uncle might post on Facebook. I could not pursue the idea any further; The follow-up question resulted in “I can’t help you with that,” which was for the best.

On the more fun and practical side, I tried to use the tool as intended, with the same prompt used in the Meta press release: “Summer is running away near me.” I pulled suggestions from some influential Facebook posts, including an AI-generated map of Puget Sound that placed the Snohomish in two outlying locations. But the basic recommendations were solid, if obvious: Whidbey Island, Mount Rainier, and hiking the Cascades.

Things got frustrating with more specific requests. She asked for things to do nearby, and suggested I swim at the community pool – noting that it would be closed over the weekend. She cited a post on the pool’s Facebook page, and the hours listed on the same page. But when I checked the source, there was no indication that the complex would be closed over the weekend – and the post in question did not appear to exist. The pool website also confirms that it is open on Saturday.

We’re about to go on a family trip to Minneapolis, and we’ll be staying downtown. I asked the AI ​​for some kid-friendly activities and got an equally confusing mix of good recommendations and hallucinations. Mita suggested an indoor mini golf venue that I had not noticed in my previous research. Nice – good. But then I suggested a café with a grass-covered play area for the kids — sounds great, except it’s in Austin. As in Texas.

I ran the query again later and got some of the same results, except for the Austin café, plus something else I didn’t see: a grill restaurant with an amazing menu, cocktails by the carafe, and a large play area for kids. It’s not exactly in the neighborhood I specified, but it’s close enough and ticks all the boxes. I think the joke is on me: Maybe AI helped me plan my summer vacation. It would have been more beneficial if he hadn’t headed to Texas first.

Photography by Alison Johnson/The Verge